Most children grasp about 300 words by the age of two, and this number expands to over 1, 000 words by the time they reach four. Although the process behind our remarkable ability to acquire language swiftly is not entirely understood, recent research in machine learning suggests that preprogrammed assumptions may not be necessary. A team of cognitive and computer scientists has successfully trained a basic artificial intelligence (AI) model to associate images with words using just 61 hours of naturalistic footage and sound, captured from the perspective of a child named Sam between 2013 and 2014. The AI was able to make sense of certain nouns, indicating that the recipe for language acquisition could be simpler than previously believed. The study, published in Science, utilized video and transcribed audio recorded by a head-mounted camera worn by Sam intermittently from six to 25 months old. Despite the limited duration of the recordings, the AI model was able to recognize the meaning of certain words. This suggests that children may not require complex, language-specific mechanisms to understand word meanings. Jessica Sullivan, an associate professor of psychology at Skidmore College, describes the study as "beautiful" and emphasizes that children's observations can provide enough information for pattern recognition and word comprehension. Furthermore, this study demonstrates that machines can learn similarly to humans. While large language models like GPT-4 are trained on massive amounts of data containing billions or even trillions of word combinations, humans require significantly less information. Bridging the gap between machine and human learning may be possible with the right kind of data. Wai Keen Vong, the lead author of the study and a computational cognitive researcher at New York University, suggests that today's models do not necessarily need as much input as they receive to make meaningful generalizations.

The study's senior author, Brenden Lake, agrees, highlighting that their research shows it is possible to train an AI model to learn words by immersing it in the experiences of a single child. To accomplish this, Lake, Vong, and their colleagues utilized a multimodal machine-learning model consisting of a vision encoder and a text encoder. These synchronized neural networks translated images and written language into the same mathematical space for interpretation by the AI. The researchers fed the model 61 hours of Sam's head-camera footage, comprised of still frames coupled with transcribed audio. Despite the seemingly messy and random nature of the dataset, which included various conversations unrelated to specific scenes or objects, both Sam and the AI model were able to understand word meanings. In multiple tests, the model accurately matched words with corresponding images and came close to the accuracy levels of two other AI models trained on significantly more language data. The study, however, has limitations. It does not prove how children acquire words, but rather indicates what is feasible for both a machine and a human. The researchers acknowledge that factors beyond mere pattern recognition likely contribute to human learning. Additionally, while the model successfully learned tens of words, there were still many it struggled to understand. It exhibited nonhuman idiosyncrasies in its recognition abilities, as it performed well with nouns for physical objects like "sand" and "car, " but poorly with words like "hand" and "room. " Human language learning encompasses much more complexity, including verbs, structure, and abstract concepts, which this study did not explore with the limited data it used. Nonetheless, this research represents a step toward a better understanding of human cognition and has the potential to benefit education and society as a whole.